xref: /freebsd/sys/dev/netmap/netmap_kern.h (revision 5686c6c38a3e1cc78804eaf5f880bda23dcf592f)
1 /*
2  * Copyright (C) 2011-2013 Matteo Landi, Luigi Rizzo. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  *   1. Redistributions of source code must retain the above copyright
8  *      notice, this list of conditions and the following disclaimer.
9  *   2. Redistributions in binary form must reproduce the above copyright
10  *      notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  */
25 
26 /*
27  * $FreeBSD$
28  *
29  * The header contains the definitions of constants and function
30  * prototypes used only in kernelspace.
31  */
32 
33 #ifndef _NET_NETMAP_KERN_H_
34 #define _NET_NETMAP_KERN_H_
35 
36 #if defined(__FreeBSD__)
37 
38 #define likely(x)	__builtin_expect(!!(x), 1)
39 #define unlikely(x)	__builtin_expect(!!(x), 0)
40 
41 #define	NM_LOCK_T	struct mtx
42 #define	NM_SELINFO_T	struct selinfo
43 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
44 #define	NM_SEND_UP(ifp, m)	((ifp)->if_input)(ifp, m)
45 
46 #elif defined (linux)
47 
48 #define	NM_LOCK_T	safe_spinlock_t	// see bsd_glue.h
49 #define	NM_SELINFO_T	wait_queue_head_t
50 #define	MBUF_LEN(m)	((m)->len)
51 #define	NM_SEND_UP(ifp, m)	netif_rx(m)
52 
53 #ifndef DEV_NETMAP
54 #define DEV_NETMAP
55 #endif
56 
57 /*
58  * IFCAP_NETMAP goes into net_device's priv_flags (if_capenable).
59  * This was 16 bits up to linux 2.6.36, so we need a 16 bit value on older
60  * platforms and tolerate the clash with IFF_DYNAMIC and IFF_BRIDGE_PORT.
61  * For the 32-bit value, 0x100000 has no clashes until at least 3.5.1
62  */
63 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,37)
64 #define IFCAP_NETMAP	0x8000
65 #else
66 #define IFCAP_NETMAP	0x100000
67 #endif
68 
69 #elif defined (__APPLE__)
70 
71 #warning apple support is incomplete.
72 #define likely(x)	__builtin_expect(!!(x), 1)
73 #define unlikely(x)	__builtin_expect(!!(x), 0)
74 #define	NM_LOCK_T	IOLock *
75 #define	NM_SELINFO_T	struct selinfo
76 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
77 #define	NM_SEND_UP(ifp, m)	((ifp)->if_input)(ifp, m)
78 
79 #else
80 
81 #error unsupported platform
82 
83 #endif /* end - platform-specific code */
84 
85 #define ND(format, ...)
86 #define D(format, ...)						\
87 	do {							\
88 		struct timeval __xxts;				\
89 		microtime(&__xxts);				\
90 		printf("%03d.%06d %s [%d] " format "\n",	\
91 		(int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec,	\
92 		__FUNCTION__, __LINE__, ##__VA_ARGS__);		\
93 	} while (0)
94 
95 /* rate limited, lps indicates how many per second */
96 #define RD(lps, format, ...)					\
97 	do {							\
98 		static int t0, __cnt;				\
99 		if (t0 != time_second) {			\
100 			t0 = time_second;			\
101 			__cnt = 0;				\
102 		}						\
103 		if (__cnt++ < lps)				\
104 			D(format, ##__VA_ARGS__);		\
105 	} while (0)
106 
107 struct netmap_adapter;
108 
109 /*
110  * private, kernel view of a ring. Keeps track of the status of
111  * a ring across system calls.
112  *
113  *	nr_hwcur	index of the next buffer to refill.
114  *			It corresponds to ring->cur - ring->reserved
115  *
116  *	nr_hwavail	the number of slots "owned" by userspace.
117  *			nr_hwavail =:= ring->avail + ring->reserved
118  *
119  * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots.
120  * This is so that, on a reset, buffers owned by userspace are not
121  * modified by the kernel. In particular:
122  * RX rings: the next empty buffer (hwcur + hwavail + hwofs) coincides with
123  * 	the next empty buffer as known by the hardware (next_to_check or so).
124  * TX rings: hwcur + hwofs coincides with next_to_send
125  *
126  * For received packets, slot->flags is set to nkr_slot_flags
127  * so we can provide a proper initial value (e.g. set NS_FORWARD
128  * when operating in 'transparent' mode).
129  */
130 struct netmap_kring {
131 	struct netmap_ring *ring;
132 	u_int nr_hwcur;
133 	int nr_hwavail;
134 	u_int nr_kflags;	/* private driver flags */
135 #define NKR_PENDINTR	0x1	// Pending interrupt.
136 	u_int nkr_num_slots;
137 
138 	uint16_t	nkr_slot_flags;	/* initial value for flags */
139 	int	nkr_hwofs;	/* offset between NIC and netmap ring */
140 	struct netmap_adapter *na;
141 	NM_SELINFO_T si;	/* poll/select wait queue */
142 	NM_LOCK_T q_lock;	/* used if no device lock available */
143 } __attribute__((__aligned__(64)));
144 
145 /*
146  * This struct extends the 'struct adapter' (or
147  * equivalent) device descriptor. It contains all fields needed to
148  * support netmap operation.
149  */
150 struct netmap_adapter {
151 	/*
152 	 * On linux we do not have a good way to tell if an interface
153 	 * is netmap-capable. So we use the following trick:
154 	 * NA(ifp) points here, and the first entry (which hopefully
155 	 * always exists and is at least 32 bits) contains a magic
156 	 * value which we can use to detect that the interface is good.
157 	 */
158 	uint32_t magic;
159 	uint32_t na_flags;	/* future place for IFCAP_NETMAP */
160 #define NAF_SKIP_INTR	1	/* use the regular interrupt handler.
161 				 * useful during initialization
162 				 */
163 	int refcount; /* number of user-space descriptors using this
164 			 interface, which is equal to the number of
165 			 struct netmap_if objs in the mapped region. */
166 	/*
167 	 * The selwakeup in the interrupt thread can use per-ring
168 	 * and/or global wait queues. We track how many clients
169 	 * of each type we have so we can optimize the drivers,
170 	 * and especially avoid huge contention on the locks.
171 	 */
172 	int na_single;	/* threads attached to a single hw queue */
173 	int na_multi;	/* threads attached to multiple hw queues */
174 
175 	int separate_locks; /* set if the interface suports different
176 			       locks for rx, tx and core. */
177 
178 	u_int num_rx_rings; /* number of adapter receive rings */
179 	u_int num_tx_rings; /* number of adapter transmit rings */
180 
181 	u_int num_tx_desc; /* number of descriptor in each queue */
182 	u_int num_rx_desc;
183 
184 	/* tx_rings and rx_rings are private but allocated
185 	 * as a contiguous chunk of memory. Each array has
186 	 * N+1 entries, for the adapter queues and for the host queue.
187 	 */
188 	struct netmap_kring *tx_rings; /* array of TX rings. */
189 	struct netmap_kring *rx_rings; /* array of RX rings. */
190 
191 	NM_SELINFO_T tx_si, rx_si;	/* global wait queues */
192 
193 	/* copy of if_qflush and if_transmit pointers, to intercept
194 	 * packets from the network stack when netmap is active.
195 	 */
196 	int     (*if_transmit)(struct ifnet *, struct mbuf *);
197 
198 	/* references to the ifnet and device routines, used by
199 	 * the generic netmap functions.
200 	 */
201 	struct ifnet *ifp; /* adapter is ifp->if_softc */
202 
203 	NM_LOCK_T core_lock;	/* used if no device lock available */
204 
205 	int (*nm_register)(struct ifnet *, int onoff);
206 	void (*nm_lock)(struct ifnet *, int what, u_int ringid);
207 	int (*nm_txsync)(struct ifnet *, u_int ring, int lock);
208 	int (*nm_rxsync)(struct ifnet *, u_int ring, int lock);
209 	/* return configuration information */
210 	int (*nm_config)(struct ifnet *, u_int *txr, u_int *txd,
211 					u_int *rxr, u_int *rxd);
212 
213 	/*
214 	 * Bridge support:
215 	 *
216 	 * bdg_port is the port number used in the bridge;
217 	 * na_bdg_refcount is a refcount used for bridge ports,
218 	 *	when it goes to 0 we can detach+free this port
219 	 *	(a bridge port is always attached if it exists;
220 	 *	it is not always registered)
221 	 */
222 	int bdg_port;
223 	int na_bdg_refcount;
224 
225 #ifdef linux
226 	struct net_device_ops nm_ndo;
227 #endif /* linux */
228 };
229 
230 /*
231  * The combination of "enable" (ifp->if_capenable & IFCAP_NETMAP)
232  * and refcount gives the status of the interface, namely:
233  *
234  *	enable	refcount	Status
235  *
236  *	FALSE	0		normal operation
237  *	FALSE	!= 0		-- (impossible)
238  *	TRUE	1		netmap mode
239  *	TRUE	0		being deleted.
240  */
241 
242 #define NETMAP_DELETING(_na)  (  ((_na)->refcount == 0) &&	\
243 	( (_na)->ifp->if_capenable & IFCAP_NETMAP) )
244 
245 /*
246  * parameters for (*nm_lock)(adapter, what, index)
247  */
248 enum {
249 	NETMAP_NO_LOCK = 0,
250 	NETMAP_CORE_LOCK, NETMAP_CORE_UNLOCK,
251 	NETMAP_TX_LOCK, NETMAP_TX_UNLOCK,
252 	NETMAP_RX_LOCK, NETMAP_RX_UNLOCK,
253 #ifdef __FreeBSD__
254 #define	NETMAP_REG_LOCK		NETMAP_CORE_LOCK
255 #define	NETMAP_REG_UNLOCK	NETMAP_CORE_UNLOCK
256 #else
257 	NETMAP_REG_LOCK, NETMAP_REG_UNLOCK
258 #endif
259 };
260 
261 /* How to handle locking support in netmap_rx_irq/netmap_tx_irq */
262 #define	NETMAP_LOCKED_ENTER	0x10000000	/* already locked on enter */
263 #define	NETMAP_LOCKED_EXIT	0x20000000	/* keep locked on exit */
264 
265 /*
266  * The following are support routines used by individual drivers to
267  * support netmap operation.
268  *
269  * netmap_attach() initializes a struct netmap_adapter, allocating the
270  * 	struct netmap_ring's and the struct selinfo.
271  *
272  * netmap_detach() frees the memory allocated by netmap_attach().
273  *
274  * netmap_start() replaces the if_transmit routine of the interface,
275  *	and is used to intercept packets coming from the stack.
276  *
277  * netmap_load_map/netmap_reload_map are helper routines to set/reset
278  *	the dmamap for a packet buffer
279  *
280  * netmap_reset() is a helper routine to be called in the driver
281  *	when reinitializing a ring.
282  */
283 int netmap_attach(struct netmap_adapter *, int);
284 void netmap_detach(struct ifnet *);
285 int netmap_start(struct ifnet *, struct mbuf *);
286 enum txrx { NR_RX = 0, NR_TX = 1 };
287 struct netmap_slot *netmap_reset(struct netmap_adapter *na,
288 	enum txrx tx, int n, u_int new_cur);
289 int netmap_ring_reinit(struct netmap_kring *);
290 
291 extern u_int netmap_buf_size;
292 #define NETMAP_BUF_SIZE	netmap_buf_size	// XXX remove
293 extern int netmap_mitigate;
294 extern int netmap_no_pendintr;
295 extern u_int netmap_total_buffers;
296 extern char *netmap_buffer_base;
297 extern int netmap_verbose;	// XXX debugging
298 enum {                                  /* verbose flags */
299 	NM_VERB_ON = 1,                 /* generic verbose */
300 	NM_VERB_HOST = 0x2,             /* verbose host stack */
301 	NM_VERB_RXSYNC = 0x10,          /* verbose on rxsync/txsync */
302 	NM_VERB_TXSYNC = 0x20,
303 	NM_VERB_RXINTR = 0x100,         /* verbose on rx/tx intr (driver) */
304 	NM_VERB_TXINTR = 0x200,
305 	NM_VERB_NIC_RXSYNC = 0x1000,    /* verbose on rx/tx intr (driver) */
306 	NM_VERB_NIC_TXSYNC = 0x2000,
307 };
308 
309 /*
310  * NA returns a pointer to the struct netmap adapter from the ifp,
311  * WNA is used to write it.
312  */
313 #ifndef WNA
314 #define	WNA(_ifp)	(_ifp)->if_pspare[0]
315 #endif
316 #define	NA(_ifp)	((struct netmap_adapter *)WNA(_ifp))
317 
318 /*
319  * Macros to determine if an interface is netmap capable or netmap enabled.
320  * See the magic field in struct netmap_adapter.
321  */
322 #ifdef __FreeBSD__
323 /*
324  * on FreeBSD just use if_capabilities and if_capenable.
325  */
326 #define NETMAP_CAPABLE(ifp)	(NA(ifp) &&		\
327 	(ifp)->if_capabilities & IFCAP_NETMAP )
328 
329 #define	NETMAP_SET_CAPABLE(ifp)				\
330 	(ifp)->if_capabilities |= IFCAP_NETMAP
331 
332 #else	/* linux */
333 
334 /*
335  * on linux:
336  * we check if NA(ifp) is set and its first element has a related
337  * magic value. The capenable is within the struct netmap_adapter.
338  */
339 #define	NETMAP_MAGIC	0x52697a7a
340 
341 #define NETMAP_CAPABLE(ifp)	(NA(ifp) &&		\
342 	((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC )
343 
344 #define	NETMAP_SET_CAPABLE(ifp)				\
345 	NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC
346 
347 #endif	/* linux */
348 
349 #ifdef __FreeBSD__
350 /* Callback invoked by the dma machinery after a successfull dmamap_load */
351 static void netmap_dmamap_cb(__unused void *arg,
352     __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error)
353 {
354 }
355 
356 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL.
357  * XXX can we do it without a callback ?
358  */
359 static inline void
360 netmap_load_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
361 {
362 	if (map)
363 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE,
364 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
365 }
366 
367 /* update the map when a buffer changes. */
368 static inline void
369 netmap_reload_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
370 {
371 	if (map) {
372 		bus_dmamap_unload(tag, map);
373 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE,
374 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
375 	}
376 }
377 #else /* linux */
378 
379 /*
380  * XXX How do we redefine these functions:
381  *
382  * on linux we need
383  *	dma_map_single(&pdev->dev, virt_addr, len, direction)
384  *	dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction
385  * The len can be implicit (on netmap it is NETMAP_BUF_SIZE)
386  * unfortunately the direction is not, so we need to change
387  * something to have a cross API
388  */
389 #define netmap_load_map(_t, _m, _b)
390 #define netmap_reload_map(_t, _m, _b)
391 #if 0
392 	struct e1000_buffer *buffer_info =  &tx_ring->buffer_info[l];
393 	/* set time_stamp *before* dma to help avoid a possible race */
394 	buffer_info->time_stamp = jiffies;
395 	buffer_info->mapped_as_page = false;
396 	buffer_info->length = len;
397 	//buffer_info->next_to_watch = l;
398 	/* reload dma map */
399 	dma_unmap_single(&adapter->pdev->dev, buffer_info->dma,
400 			NETMAP_BUF_SIZE, DMA_TO_DEVICE);
401 	buffer_info->dma = dma_map_single(&adapter->pdev->dev,
402 			addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE);
403 
404 	if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) {
405 		D("dma mapping error");
406 		/* goto dma_error; See e1000_put_txbuf() */
407 		/* XXX reset */
408 	}
409 	tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX
410 
411 #endif
412 
413 /*
414  * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction.
415  */
416 #define bus_dmamap_sync(_a, _b, _c)
417 
418 #endif /* linux */
419 
420 /*
421  * functions to map NIC to KRING indexes (n2k) and vice versa (k2n)
422  */
423 static inline int
424 netmap_idx_n2k(struct netmap_kring *kr, int idx)
425 {
426 	int n = kr->nkr_num_slots;
427 	idx += kr->nkr_hwofs;
428 	if (idx < 0)
429 		return idx + n;
430 	else if (idx < n)
431 		return idx;
432 	else
433 		return idx - n;
434 }
435 
436 
437 static inline int
438 netmap_idx_k2n(struct netmap_kring *kr, int idx)
439 {
440 	int n = kr->nkr_num_slots;
441 	idx -= kr->nkr_hwofs;
442 	if (idx < 0)
443 		return idx + n;
444 	else if (idx < n)
445 		return idx;
446 	else
447 		return idx - n;
448 }
449 
450 
451 /* Entries of the look-up table. */
452 struct lut_entry {
453 	void *vaddr;		/* virtual address. */
454 	vm_paddr_t paddr;	/* physical address. */
455 };
456 
457 struct netmap_obj_pool;
458 extern struct lut_entry *netmap_buffer_lut;
459 #define NMB_VA(i)	(netmap_buffer_lut[i].vaddr)
460 #define NMB_PA(i)	(netmap_buffer_lut[i].paddr)
461 
462 /*
463  * NMB return the virtual address of a buffer (buffer 0 on bad index)
464  * PNMB also fills the physical address
465  */
466 static inline void *
467 NMB(struct netmap_slot *slot)
468 {
469 	uint32_t i = slot->buf_idx;
470 	return (unlikely(i >= netmap_total_buffers)) ?  NMB_VA(0) : NMB_VA(i);
471 }
472 
473 static inline void *
474 PNMB(struct netmap_slot *slot, uint64_t *pp)
475 {
476 	uint32_t i = slot->buf_idx;
477 	void *ret = (i >= netmap_total_buffers) ? NMB_VA(0) : NMB_VA(i);
478 
479 	*pp = (i >= netmap_total_buffers) ? NMB_PA(0) : NMB_PA(i);
480 	return ret;
481 }
482 
483 /* default functions to handle rx/tx interrupts */
484 int netmap_rx_irq(struct ifnet *, int, int *);
485 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL)
486 
487 #endif /* _NET_NETMAP_KERN_H_ */
488